arXiv:1706.03989v2 [math.CA] 22 Jun 2018 euto h osdrto,w eieafruawihrepresents which formula a derive we consideration, the of P result a qain()b h hoyo cwrzdsrbtos nti pape this conve In the prove th distributions. general to show Schwartz for not hard of (2) wh does is theory equation certain [8] It formal the not in by is discussion not. (2) it The or Equation Actuary, approximation. convergent Laplace sence. is the naive (2) with a of in side equation hand formal a only is itiuino ahvariable each of distribution P Let Introduction 1 o vlaino h est function. density the of evaluation for admvrals h u fthe of statistics sum chi-squared the the variables, of generalization random a As function. density hsppr edsustesmo the of sum the discuss we paper, this ucincnb rte ntefr of form the in written be can function et n eaeitrse nnmrclaayi ftepoaiiyd probability the of analysis numerical e in be interested can are quantity we the and on investigation test, Detailed statistics. in quantity ttsis h u fthe of sum The statistics. ntecs where case the In where egtdsum weighted h egtdsmi icse n[]ad[] n[] auo au n Ta where and Oaku, case Marumo, the [8], in In (1) [7]. integral and of [1] in analysis of discussed analysis is Numerical sum weighted statistics. the chi-squared the of generalization Catr6o 3) hydrvdasse fdffrnileutosfo equations differential of system a derived They of [3]). of 6 ,Chapter k n k n P =1 =1 nodrt eietesse fdffrnileutosfrtepro the for equations differential of system the derive to order In nitga oml o h oee u fteindependent, the of sum powered the for formula integral An { X k n oml sbsdo h neso oml,adw tlz su hyperfunctions. a of utilize point we view and evaluatio the formula, for in inversion applied formula the our be on to based is expected formula is represen formula we integral paper, converge Our absolutely this one-dimensional In an by variable random distribution. chi-squared the of =1 X X ϕ k } k k ( r 3 h itiuino h u of sum the of distribution The X t k auo au n aeuauiietefloigivrinformula: inversion following the utilize Takemura and Oaku, Marumo, . dnial n omlydsrbtdrno variables random distributed normally and identically := ) sa n-iesoa boueycnegn nerl hsformula This integral. convergent absolutely one-dimensional an as ∈ k 3 N ynmrclysligtesse fdffrnileutos hyev they equations, differential of system the solving numerically By . easqec fteidpnet dnial itiue admva random distributed identically independent, the of sequence a be F P E ( k n c =1 := ) exp r w ,ti nerleul otecmltv itiuinfnto fth of function distribution cumulative the to equals integral this 2, = k P itX X

k 2 where , k X k 3 =1 n r  t power -th X X ( k r i k r stesadr omldsrbto.Fxapstv integer positive a Fix distribution. normal standard the is := ≥ c < w r rmtevepito umto ehd n yefntos As hyperfunctions. and methods summation of viewpoint the from 3 k t oe ( power -th √ ! r sapstv elnme,o h qae omlvralsi anothe is variables normal squared the of number, real positive a is r f t oe fsadr omlrno aibe sageneraliz a is variables random normal standard of power -th − t oe fsadr omlrno aibe.Ti distribution This variables. random normal standard of power -th n ( P = x )i h hrceitcfnto of function characteristic the is 1) iesoa integral: dimensional = ) r k n Z =1 .Terapoc a h oooi rdetmto ([9] method gradient holonomic the was approach Their 3. = ai Koyama Tamio P k n 2 X =1 1 π Abstract k r r x Z k r sa eeaiaino h h-qae ttsis The statistics. chi-squared the of generalization an is −∞ ≥

h uuaiedsrbto ucinof function distribution cumulative the − ,w osdrajsicto fthe of justification a consider we r, pce ob ple oHypothesis to applied be to xpected niyfnto ftequantity. the of function ensity h rbblt est ucinof function density probability the 2 1 gneo h nerlb utilizing by integral the of rgence h rbblt est function density probability the r aiiydniyfunction density bability k X =1 n ovrec f()adjustify and (2) of convergence e euadsusdtenumerical the discussed kemura te h nerli h right- the in integral the ether x X k 2 ! k 3 oeta qain(2) Equation that Note . sepce ob applied be to expected is dx ibe,adspoethe suppose and riables, 1 laeteprobability the aluate dx . . . n chi-squared e r ation and f ( n x In . of ) (1) (2) r Before starting regorous discussion, let us give an intuitive discussion on the formal equation (2). Here, we consider the case where ϕ(t) is the characteristic function of a general random varialbe X. Note that the distribution of X can be the exponential distribution, the binomial distribution or other distribution. Dividing the domain of the formal integral in the right-hand side of (2), it can be written as 0 1 itx n 1 ∞ itx n e− ϕ(t) dt + e− ϕ(t) dt. (3) 2π 2π 0 Z−∞ Z Note that the integral in the first term converges when x > 0 holds, the second term converges when x < 0. Hence, the first and the second term define holomorphicℑ functions on the upper half plane andℑ the lower half plane respectively. Since these holomorphic functions has analytic continuation, the expression (3) make a sense for x R 0 if x is not a singular point of the holomorphic functions. This intuitive explanation can be justified∈ \{ by} a summation methods in the case where random variable X is the sum of r-th power of standard normal random variables. For general random variables, this intuitive explanation is justified by theory of Fourier hyperfunctions. Theory of hyperfunction introduced in [11],[6] has many applications to linear partial differential equations (see, e.g., [2]). Applications of hyperfunctions to numerical analysis are discussed in [4] and [10]. Our discussion in this paper is an application of the theory of hyperfunctions to statistics. The organization of this paper is as follows. In Section 2, we review the inversion formulae in the probability theory and the Fourier analysis, and we justify the intuitive discussion in Section 1 by a summation method. In Section 3, we review the theory of hyperfunctions of a single variable based on [11] and [5], and we justify the intuitive discussion in Section 1 from the view point of hyperfunctions. In Section 4, we apply the inversion formula given in Section 2 to the sum of r-th power of standard normal random variables. We also calculate the limit in the boundary-value representation.

2 Inversion Formula

In this section, we review the inversion formulae in the probability theory and the Fourier analysis. In the probability theory, the following L´evy’s inversion formula is most famous. Theorem 1 (L´evy’s inversion formula). Let X be a random variable and ϕ(t) be the characteristic function of X. Then, for a

R ita itb 1 e− e− 1 1 lim − ϕ(t)dt = P (X = a)+ P (a

1 (0

2 If f is a slowly increasing function, Equation (4) can be justified as an equation of Schwartz distributions. The following proposition is a justification of the intuitive discussion in Section 1 by a summation method.

Proposition 1. Let f be a probability density function and ϕ(t) := ∞ eityf(y)dy be the characteristic function of f. For any continuous point x of f, the equation −∞ R 0 1 i(x+iε)t ∞ i(x iε)t f(x)= lim e− ϕ(t)dt + e− − ϕ(t)dt 2π ε +0 0 → Z−∞ Z  holds.

We utilize a fact in [5, p28, Note 1.3] with a small change. Lemma 1. Let K(x) and ϕ(x) be real-valued functions on R. Suppose that K(x) is absolutely integrable and ∞ K(x)dx =1 holds, and that ϕ(x) is continuous ans bounded. For a positive number ε> 0, put −∞ R 1 t x ϕε(x) := K − ϕ(t)dt. ε R ε Z  

Then, for any x R, we have ϕ(x) = limε +0 ϕε(x). ∈ → Proof. Since ϕε(x) is decomposed as

∞ x t dt ∞ x t dt ϕ (x)= ϕ(x) K − (ϕ(x) ϕ(t)) K − ε ε ε − − ε ε Z−∞   Z−∞   ∞ x t dt = ϕ(x) (ϕ(x) ϕ(t)) K − , − − ε ε Z−∞   it is enough to show that the second term converges to zero. We decompose the integral domain into x t ε and x t <ε. On the first domain, we have | − |≥ | − | x t dt x t dt (ϕ(x) ϕ(t)) K − 2 sup ϕ(x) K − x t ε − ε ε ≤ x | | x t ε ε ε Z| − |≥     Z| − |≥  

= 2 sup ϕ(x) K (s) ds. x | | s 1/√ε | |   Z| |≥ Note that sup ϕ(x) is finite since ϕ(x) is bounded. Since K(x) is absolutely integrable, K (s) ds x | | s 1/√ε | | converges to 0 as ε +0. Consequenty, we have | |≥ → R x t dt (ϕ(x) ϕ(t)) K − 0 (ε +0). x t ε − ε ε → → Z| − |≥  

On the second domain, we have

x t dt (ϕ(x) ϕ(t)) K − sup ϕ(x) ϕ(t) K (s) ds x t <ε − ε ε ≤ x t <ε | − |! s 1/√ε | | Z| − |   | − | Z| |≥

∞ sup ϕ(x) ϕ(t) K (s) ds. ≤ x t <ε | − | | | | − | ! Z−∞

Since ϕ(x) is continuous, sup x t <ε ϕ(x) ϕ(t) goes to zero as ε +0. Consequently, we have | − | | − | → x t dt (ϕ(x) ϕ(t)) K − 0 (ε +0). x t <ε − ε ε → → Z| − |  

Therefore, we have ϕ(x) = limε +0 ϕε(x). →

3 Proof of Proposition 1. Fix the continuous point x of f. Decompose f into a sum of two continuous functions f1 and f2 where f1 is compact and f2 equals to zero on a neighborhood of x. Put ity ϕj (t) := ∞ e fj(y)dy (j =1, 2). By the Fubini’s theorem, we have −∞ R 0 1 i(x+iε)t ∞ i(x iε)t e− ϕ1(t)dt + e− − ϕ1(t)dt 2π 0 Z−∞ Z  0 1 ∞ i(y x iε)t ∞ ∞ i(y x+iε)t = f1(y) e − − dtdy + f1(y) e − dtdy 2π 0 Z−∞ Z−∞ Z−∞ Z  ∞ 1 ε = f (y) dy 1 π (y x)2 + ε2 Z−∞ − 1 Apply Lemma 1 with K(x)= π(x2+1) , then we have

∞ 1 ε f (y) dy f (x) (ε +0). 1 π (y x)2 + ε2 → 1 → Z−∞ − Analogously, we have

0 1 i(x+iε)t ∞ i(x iε)t 1 ∞ 1 1 e− ϕ2(t)dt + e− − ϕ2(t)dt = f2(y) dy 2π 0 2π i(y x iε) − i(y x + iε) Z−∞ Z  Z−∞  − − −  Since f2 equals to zero on a neighborhood of x, for sufficiently small δ > 0, the above integral equals to 1 1 1 f2(y) dy 2π y x >δ i(y x iε) − i(y x + iε) Z| − |  − − −  By the Lebesgue’s dominated convergence theorem, this integral converges to 0 as ε +0. Hence, we have → 0 1 i(x+iε)t ∞ i(x iε)t lim e− ϕ(t)dt + e− − ϕ(t)dt 2π ε +0 0 → Z−∞ Z  2 0 1 i(x+iε)t ∞ i(x iε)t = lim e− ϕj (t)dt + e− − ϕj (t)dt = f1(x)= f(x). 2π ε +0 j=1 → 0 X Z−∞ Z 

3 Perspective of Hyperfunction

In this section, we briefly review the theory of hyperfunctions and justify the intuitive discussion in Section 1 from the view point of hyperfunctions. In the first, we review the theory of hyperfunctions of a single variable based on [11] and [5]. Let be the of holomorphic functions on the complex plane C. The sheaf of the hyperfunctions ofO a 1 B single variable is defined as the 1-th derived sheaf R( ) of . The global section (R) is the inductive limit lim (U R)/O(U) with respect to the familyH O of complexO neighborhoods UB R. The elements U R O \ ⊃ of (R−→) are⊃ called hyperfunctions on R. Any hyperfunction on R can be represented as a equivalent classB [F (z)] with a representative F (z) (U R). ∈ O \ Let L1,loc(R) be the spaces of locally integrable functions on R. There is a natural embedding ι from L1,loc(R) to (R). When an integrable function f L1(R) satisfies some condition, we can take a representative F (zB) of ι(f) as a on∈C R. \ Proposition 2 (M. Sato). Let f(x) be an integrable function on R such that the integral ∞ (x2 + 1 −∞ 1)− f(x)dx is finite. Take a constant α C R, and put ∈ − R 1 ∞ 1 1 F (z)= f(x) dx (z C R). 2πi x z − x α ∈ \ Z−∞  − −  Then we have ι(f) = [F ].

4 Proof. See, [11]. We call a measure µ on R to be locally integrable when µ(K) is finite for any compact set K R. ⊂ We denote by Mloc the space of the locally integrable measures on R. Analogous to the case of L1,loc(R), there is a natural embedding ι from M to (R), and the following proposition holds: loc B 2 1 Proposition 3 (M. Sato). Let µ be a measure on R such that the integral R(x + 1)− µ(dx) is finite. Take a constant α C R, and put ∈ − R 1 1 1 F (z)= µ(dx) (z C R) 2πi R x z − x α ∈ \ Z  − −  Then we have ι(µ) = [F ].

Let I be a neighborhood of 0 R, and put I∗ := I 0 . A function F (z) holomorphic on a tubular ∈ \{ } domain R + iI∗ := z C z I∗ is said to be slowly increasing if for any compact subset K I∗ and any ε> 0, there exists{ ∈C >|ℑ0 such∈ } that ⊂ ε z F (z) Ce |ℜ | | |≤ uniformly for z R + iK. A hyperfunction f on R is said to be slowly increasing if we can take a slowly increasing function∈ F (z) as its representative, i.e., there exists a slowly increasing function F (z) such that f = [F (z)]. Put

z z e e− χ+(z) := z , χ (z) := z . 1+ e − 1+ e− The following definition is a specialization into the case of a single variable of the general definition in [5].

Definition 1. Let a slowly increasing function F (z) on a tubular domain R + iI∗ be a representative of a slowly increasing hyperfunction f (R). Take a positive number y I∗. We define the Fourier transform of f = [F (z)] as f = [G(w)]∈ where B ∈ F iwz iwz e− χ (z)F (z)dz e− χ (z)F (z)dz ( w> 0) z=+y − z= y − G(w)= ℑ iwz − ℑ − iwz ℑ (R z=+y e− χ+(z)F (z)dz +R z= y e− χ+(z)F (z)dz ( w< 0). − ℑ ℑ − ℑ Here, for y R and a functionR ϕ(z) with a complex variableR z, we denote the integral ϕ(x + iy)dx by ∈ R z=y ϕ(z)dz. ℑ R R In the second, we give and show an equation corresponding to the formal equation (2). Let X be a itX random variable. The distribution µX of X and the characteristic function ϕX (t) = E e of X can be regarded as hyperfunctions. We denote them by ιµ and ιϕ respectively. The following equation x X  corresponds to the formal equation (2).

(ιϕ )=2πιµ . (6) F X X Since the characteristic function can be regarded as the inverse Fourier transformation of the distribution under a suitable condition, the equation (6) is an inversion formula. We calculate defining functions of ιµX and ιϕX . Let α C R be a constant. Since µX is a probability measure on R, we have ∈ \ 1 1 2 µX (dx) 2 µX (dx) µX (dx)=1 < R (x + 1) ≤ R x +1 ≤ R ∞ Z Z Z | | and 1 1 1 1 µX (dx) µX (dx) µX (dx)= < . R x α ≤ R x α ≤ R α α ∞ Z Z Z − | − | |ℑ | |ℑ | By Proposition 3, we have

1 1 1 1 1 ιµX = µX (dx) = µX (dx) . (7) 2πi R x z − x α 2πi R x z  Z  − −    Z − 

5 2 1 Since the estimation ∞ ϕX (t)(t + 1)− dt < holds, we can apply Proposition 2 to ϕX (t). In fact, −∞ ∞ we have R

∞ ϕX (t) ∞ ϕX (t) ∞ 1 ∞ 2 dt 2 dt 2 dt = [arctan(t)] = π. t +1 ≤ t +1 ≤ t +1 −∞ Z−∞ Z−∞ Z−∞

Hence, we have 1 ∞ 1 1 ιϕ = ϕ (t) dt . X 2πi X t w − t α  Z−∞  − −   We calculate the Fourier transformation of ιϕX .

Lemma 2. For a random variable X, the hyperfunction ιϕX corresponding to the characteristic function ϕX of X is slowly increasing. Proof. For simplicity, we assume α = √ 1 without loss of generality. The absolute value of the repre- − sentative of ιϕX satisfies the inequality

1 ∞ 1 1 w α ∞ dt ϕ (t) dt | − | . (8) 2πi X t w − t α ≤ 2π t w t α Z−∞   Z−∞ − − | − || − | Put c := w , then the inequalities |ℜ | (c+1) c 1 ∞ dt ∞ dt − dt − − dt 2 =1, 2 =1, c+1 t w t α ≤ c+1 (t c) t w t α ≤ (t + c) Z | − || − | Z − Z−∞ | − || − | Z−∞ c+1 dt 2(c + 1) , c 1 t w t α ≤ w Z− − | − || − | |ℑ | imply that the right hand side of (8) is bounded above by

w α w +1 | − | 1+ |ℜ | . π w  |ℑ | 

Hence, ιϕX is a slowly increasing hyperfunction.

In order to calculate an explicit form of a defining function of the Fourier transformation of ιϕX , we decompose the function ϕX (t) as

ϕ (t)=(1 H(t))ϕ (t)+ H(t)ϕ (t) X − X X where H(t) is the Heaviside function. Embedding the both sides of the above equation into the global section B(R) of the sheaf of hyperfunctions, we have

ιϕ = ι((1 H)ϕ )+ ι(Hϕ ) X − X X By Proposition 2, defining functions of ι((1 H)ϕ ) and ι(Hϕ ) are given by − X X 1 0 1 1 ϕ1(w) := ϕX (t) dt (w C R 0), 2πi t w − t α ∈ − ≤ Z−∞  − −  1 ∞ 1 1 ϕ2(w) := ϕX (t) dt (w C R 0) 2πi t w − t α ∈ − ≥ Z0  − −  respectively. Then we have ιϕX = [ϕ1(w)] + [ϕ2(w)]. Analogous to Lemma 2, we can show that both [ϕ1(w)] and [ϕ2(w)] are slowly increasing hyperfunctions. (0+) For a function F (z) with a complex variable , we denote by + F (z)dz the integral along the path (0+) ∞ described in Figure 1. We also denote by F (z)dz the integralR along the path described in Figure 2. We denote by H the upper half-plane −∞z C z > 0 . R{ ∈ |ℑ }

6 O O

(0+) Figure 1: Contour of the path of + (0+) ∞ Figure 2: Contour of the path of R −∞ R Lemma 3. The holomorphic functions

(0+) izw e− ϕ1(w)dw (z H), ψ1(z) := − −∞ ∈ (0 (z H), R ∈− 0 (z H), ψ (z) := ∈ 2 (0+) izw ( + e− ϕ2(w)dw (z H), ∞ ∈− R are defining functions of the Fourier transformations of [ϕ1(w)] and [ϕ2(w)] respectively, i.e., we have [ϕ (w)] = [ψ (z)] (k =1, 2). F k k Proof. By Definition 1, a defining function of [ϕ (w)] is given by F 1 izw izw e− χ (w)ϕ1(w)dw e− χ (w)ϕ1(w)dw ( z > 0) ˜ w=+y − w= y − ψ1(z)= ℑ izw − ℑ − izw ℑ (R w=+y e− χ+(w)ϕ1(w)dw +R w= y e− χ+(w)ϕ1(w)dw ( z < 0). − ℑ ℑ − ℑ R R Since ϕ1 is holomorphic on C R 0, we have − ≤ (0+) izw e− χ (w)ϕ1(w)dw ( z > 0) ψ˜ (z)= − − ℑ 1 (0+)−∞ izw ( R e− χ+(w)ϕ1(w)dw ( z < 0). −∞ ℑ

(0+) izw R Note that the integral e− χ+(w)ϕ1(w)dw defines a holomorphic function on a neighborhood of R. Hence, −∞ R (0+) izw ψ (z)= ψ˜ (z) e− χ (w)ϕ (w)dw 1 1 − + 1 Z−∞ is a defining function of [ϕ1(w)] also. We can show [ϕ (wF)] = [ψ (z)] analogously. F 2 2 Lemma 4. Put 0 izt e− ϕX (t)dt (z H), ψ(z)= −∞ izt ∈ ∞ e− ϕ (t)dt (z H). (R− 0 X ∈− Then ψ(z) is a defining function of ιϕ . R F X

7 Proof. It is enough to calculate the right hand side of the equation ιϕX = [ψ1 + ψ2]. For z H, we have F ∈− (0+) izw 1 ∞ 1 1 ψ2(z)= e− ϕX (t) dtdw + 2πi 0 t w − t α Z ∞ Z  − −  (0+) 1 ∞ izw 1 1 = e− ϕX (t) dtdw 2πi + 0 t w − t α Z ∞ Z  − −  By the assumption z H, we have ∈− (0+) ∞ izw 1 1 e− ϕX (t) dtdw < . + 0 t w − t α ∞ Z ∞ Z   − − By the Fubini-Tonelli theorem and Cauchy’s integral formula, ψ (z ) equals to 2 (0+) 1 ∞ izw 1 1 e− ϕX (t) dwdt 2πi 0 + t w − t α Z Z ∞  − −  (0+) 1 ∞ izw 1 1 ∞ izt = ϕX (t) e− + dwdt = e− ϕX (t)dt. −2πi 0 + w t t α − 0 Z Z ∞  − −  Z 0 izt For z H, we can show ψ1(z)= e− ϕX (t)dt analogously. ∈ −∞ Lemma 5. For z C R, the followingR equation holds: ∈ \ 1 1 ψ(z)= µX (dx). (9) i R x z Z − Proof. For z H, we have ∈− ∞ izt ∞ i(z x)t e− ϕX (t)dt = e− − µX (dx)dt (10) − − R Z0 Z0 Z by the definition of the characteristic function. Since z is negative, we have ℑ ∞ i(z x)t ∞ zt 1 e− − µX (dx)dt = eℑ dt < . 0 R 0 ≤− z ∞ Z Z Z ℑ

By the Fubini-Tonelli theorem, the right hand side of (10) equals to ∞ i(z x)t 1 1 1 e− − dtµX (dx)= µX (dx)= µX (dx). − R R i(z x) i R x z Z Z0 Z − − Z − For z H, we can show the equation (9) analogously. ∈ Theorem 2. For a random variable X, equation (6) holds. Proof. By Lemma 5 and equation (7), we have 1 1 ιϕX = [ψ(z)] = µX (dx) =2πιµX . F i R x z  Z − 

Other Proof of Proposition 1. Let X be a random variable whose probability density function is f(x). Then, ϕ(t) equals to the characteristic function of X. By Lemma 2 and Lemma 4, the hyperfunction ιϕ corresponding to ϕ(t) is a Fourier hyperfunction, and a defining function of ιϕ is F 0 izt e− ϕ(t)dt (z H), ψ(z) := −∞ izt ∈ ∞ e− ϕ(t)dt (z H). (−R 0 ∈− 1 By Theorem 2, the corresponding hyperfunctionR ιf of f equals to (2π)− ιϕ. For any continuous point x of f, the boundary-value representation [5, Theorem 1.3.12] implies F 0 1 i(x+iy)t ∞ i(x iy)t f(x)= lim e− ϕ(t)dt + e− − ϕ(t)dt 2π y +0 0 → Z−∞ Z 

8 4 Sum of r-th Power of the Normal Variables

Fix a positive integer r 3. Let X1,...,Xn be the independent, identically and normally distributed ≥ n r r random variables, and put X := k=1 Xk . Let ϕ(t) be the characteristic function of X1 , i.e., put

P r r 1 2 itX1 itx 1 x ϕ(t)= E e = e e− 2 dx. (11) R √2π   Z n Then, the characteristic function ϕX of the powered sum X equals to ϕ(t) . By Proposition 1, we have

0 1 i(x+iy)t n ∞ i(x iy)t n fX (x)= lim e− ϕ(t) dt + e− − ϕ(t) dt (12) 2π y +0 0 → Z−∞ Z  for any continuous point x R of fX . In order to calculate the∈ limit in the right hand side of (12), we discuss on the analytic continuations of the each term of (3). In the first, we consider the characteristic function in (11). The characteristic function can be decom- posed as 0 1 ∞ 1 1 1 ϕ(t)= exp itxr x2 dx + exp itxr x2 dx. √2π 0 − 2 √2π − 2 Z   Z−∞   We are interested in the analytic continuation of each term in the right hand side. In order to calculate them, we consider the following general form of integral:

1 r 1 2 ϕθ(w) := exp iwz z dz (w C), (13) √2π − 2 ∈ Zγθ   where the integral path γ with parameter θ R is defined by θ ∈ γ (t) := teiθ (0

Note that we have ϕ(t)= ϕ0(t) ϕπ(t) for t R. For z C, put − ∈ ∈ zH := zw w H . { | ∈ } Examining the convergence region of the integral (13) for given parameter θ R, we obtain the following lemma: ∈

iθr Lemma 6. Suppose that d 3. For given parameter θ R, the integral (13) converges on e− H. ≥ ∈ iθr Proof. Take any point w e− H. By the straight forward calculation, we have ∈

1 ∞ iθr r 1 2iθ 2 iθ ϕθ(w) exp iwe t e t e dt | |≤ √2π − 2 Z0  

1 ∞ iθr r cos(2θ) 2 = exp (we )t t dt √2π −ℑ − 2 Z0   Since assumption weiθr H implies (weiθr) > 0, there exist t > 1 and ε> 0 such that ∈ ℑ 0

iθr cos(2θ) (we ) r 2 < ε −ℑ − 2t − − holds for any t>t0. Hence, we have

∞ cos(2θ) exp (weiθr)tr t2 dt −ℑ − 2 Zt0   ∞ r ∞ r 1 r exp( εt0) exp( εt ) dt r t − exp( εt ) dt = − . ≤ − ≤ · − ε Zt0 Zt0 iθr Consequently, the integral (13) converges on e− H.

9 iθr Lemma 7. Suppose that d 3. If cos2θ > 0 holds, then the integral (13) converges on e− H¯ . Here, H¯ denotes the closure of H.≥

iθr iθr Proof. For any point w e− H¯ , we have (we ) 0, This implies ∈ ℑ ≥

1 ∞ iθr r cos(2θ) 2 ϕθ(w) exp (we )t t dt | |≤ √2π −ℑ − 2 Z0   1 ∞ cos(2θ) exp t2 dt < ≤ √2π − 2 ∞ Z0  

Example 1. Let r = 3. Figures 3 and 4 show the contour of the integral path γθ and the region exp( iθr) H for θ = 0 respectively. Figures 5 and 6 show them for θ = π/9. −

γ

Figure 3: The integral path for θ =0 Figure 4: The convergence region for θ =0

′ iθr iθ r Lemma 8. If θ θ′ < π/d, then ϕ (w)= ϕ ′ (w) holds for w e− H e− H. | − | θ θ ∈ ∩ Proof. Put

′ ′′ iθ iθ iθ γ (t) := te (0 t R), γ (t) := te (0 t R), γ (θ′′) := Re (θ θ′′ θ′). 1,R ≤ ≤ 2,R ≤ ≤ 3,R ≤ ≤ By the Cauchy’s integral formula, we have

Φdz Φdz + Φdz =0. (15) − Zγ1,R Zγ2,R Zγ3,R 1/2 r 2 Here, we put Φ = (2π)− exp iwz z /2 . By some calculations, we have that there exist M > 0 and ε> 0 such that −  Φdz MR exp( εRr) . γ3,R ≤ − Z

Taking the limit of (15) as R approaches inifinity, we have ϕ (z) ϕ ′ (z)=0. θ − θ In the second, we disscuss on the analytic continuation of the each term of (3). Let us consider the following integral: iwz n ψ(z; θ,θ′, φ) := e− (ϕ (w) ϕ ′ (w)) dw. (16) θ − θ Zγφ where θ,θ′, φ R, z C, and the integral path γ is defined by (14). ∈ ∈ φ

10 γ θd

θ

Figure 5: The integral path for θ = π/9 Figure 6: The convergence region for θ = π/9

′ Lemma 9. Suppose (ei(φ+θr)) 0 and (ei(φ+θ r)) 0. Then the integral (16) converges for z exp( iφ)H. ℑ ≥ ℑ ≥ ∈ − − Proof. By the binomial expansion, the integral (16) equals to

n n n k iwz k n k ( 1) − e− ϕθ(w) ϕθ′ (w) − dw. k − γφ kX=0   Z With some calculation, the each term of the above expression can be written as

1 ∞ iwz k ′ n k e− ϕθ(w) ϕθ (w) − dw = n/2 exp( h1(z,t)s + h2(t)) dtds, (2π) Rn − Zγφ Z0 Z >0 where we put t := (t1,...,tn)⊤, dt := dt1 ...dtn,

k n k n ′ ′ iφ iθr r iθ r r 1 2iθ 2 2iθ 2 h1(z,t)= ie z e tℓ e tℓ , h2(t)= e tℓ + e tℓ . − − ! −2 ! Xℓ=1 ℓ=Xk+1 Xℓ=1 ℓ=Xk+1 By the Fubini-Tonelli theorem, it is enough to consider the convergence of the following integral:

∞ exp( h1(z,t)s + h2(t)) dsdt. (17) Rn | − | Z >0 Z0 For z exp( iφ)H, we have (izeiφ) > 0. This inequality and the assumption ( iei(φ+θr)) 0 and ′ ( ie∈−i(φ+θ r)) − 0 imply h (z,tℜ ) > 0. Hence, the integral (17) equals to ℜ − ≥ ℜ − ≥ ℜ 1 ∞ exp(h2(t)) exp(h2(t)) exp( s h1(z,t)) dsdt = | |dt. (18) Rn | | − ℜ Rn h1(z,t) Z >0 Z0 Z >0 ℜ By the assumptions, the integral (18) is bounded from above by

k n k 1 1 ∞ 1 2 ∞ 1 ′ 2 − 2 cos(2θ)t 2 cos(2θ )t iφ exp(h2(t)) dt = iφ e− dt e− dt (ize ) Rn | | (ize ) ℜ Z >0 ℜ Z0  Z0  k/2 (n k)/2 1 π π − = < . (izeiφ) 2 cos(θ) 2 cos(θ ) ∞ ℜ    ′  Therefore, the integral (16) converges.

11 ′ Lemma 10. Suppose (ei(φk+θr)) 0 and (ei(φk+θ r)) 0 for k =1, 2. The equation ℑ ≥ ℑ ≥

ψ(z; θ,θ′, φ1)= ψ(z; θ,θ′, φ2) holds for z ( exp( iφ )) ( exp( iφ )). ∈ − − 1 ∩ − − 2 Proof. Put

γ (t) := teiφ1 (0 t R), γ (t) := teiφ2 (0 t R), γ (φ) := Reiφ (φ φ φ ). 1,R ≤ ≤ 2,R ≤ ≤ 3,R 1 ≤ ≤ 2 By the Cauchy’s integral formula, we have

Φdw Φdw + Φdw =0, (19) − Zγ1,R Zγ2,R Zγ3,R iwz n where we put Φ = e− (ϕ (w) ϕ ′ (w)) . By some calculations, we have θ − θ

Φdw MR exp( R) γ3,R ≤ − Z

for sufficiently large M > 0. Taking the limit of (19) as R approaches inifinity, we have ψ(z; θ,θ′, φ ) 1 − ψ(z; θ,θ′, φ2)=0.

Theorem 3. Let ε> 0 be a sufficiently small positive number and x be a continuous point of fX . When r is odd, we have

ψ(x; ε, π, εr) ψ(x; ε,π,π + εr) (x> 0), fX (x)= − − − ψ(x;0, π ε,εr) ψ(x;0, π + ε, π εr) (x< 0). ( − − − When r is even, we have

ψ(x; ε, π + ε, εr) ψ(x; ε, π ε, π + εr) (x> 0), fX (x)= − − − − ψ(x;0,π,εr) ψ(x;0,π,π εr) (x< 0). ( − − Proof. When r is odd and x> 0, we have 1 fX (x)= lim (ψ(x iy;0, π, 0) ψ(x + iy;0,π,π)) y +0 2π → − − 1 = lim (ψ(x iy; ε, π, 0) ψ(x + iy;0, π + ε, π)) y +0 2π → − − 1 = lim (ψ(x iy; ε, π, εr) ψ(x + iy;0, π + ε, π + εr)) y +0 2π → − − − 1 = (ψ(x; ε, π, εr) ψ(x;0, π + ε, π + εr)) 2π − − When r is odd and x< 0, we have 1 fX (x)= lim (ψ(x iy;0, π, 0) ψ(x + iy;0,π,π)) 2π y +0 → − − 1 = lim (ψ(x iy; ε, π, 0) ψ(x + iy;0, π ε, π)) 2π y +0 → − − − − 1 = lim (ψ(x iy; ε,π,εr) ψ(x + iy;0, π ε, π εr)) y +0 2π → − − − − − 1 = (ψ(x; ε,π,εr) ψ(x;0, π ε, π εr)) 2π − − − −

12 When r is even and x> 0, we have 1 fX (x)= lim (ψ(x iy;0, π, 0) ψ(x + iy;0,π,π)) y +0 2π → − − 1 = lim (ψ(x iy; ε, π + ε, 0) ψ(x + iy;0, π + ε, π)) y +0 2π → − − 1 = lim (ψ(x iy; ε, π + ε, εr) ψ(x + iy;0, π + ε, π + εr)) 2π y +0 → − − − 1 = (ψ(x; ε, π + ε, εr) ψ(x;0, π + ε, π + εr)) 2π − − When r is odd and x< 0, we have 1 fX (x)= lim (ψ(x iy;0, π, 0) ψ(x + iy;0,π,π)) y +0 2π → − − 1 = lim (ψ(x iy; ε, π, 0) ψ(x + iy;0, π ε, π)) 2π y +0 → − − − − 1 = lim (ψ(x iy; ε,π,εr) ψ(x + iy;0, π ε, π εr)) 2π y +0 → − − − − − 1 = (ψ(x; ε,π,εr) ψ(x;0, π ε, π εr)) 2π − − − −

Remark 1. The all of integral representations in Theorem 3 make sense as Lebesgue integral. As we have shown in Lemma 9, when we write the integral in Theorem 3 into an integral on R, the integrand is a Lebesgue integrable function. Acknowledgements: We are grateful to K. Yajima, T. Oshima, and S.-J. Matsubara-Heo for valuable comments to Proposition 1. This work was supported by MEXT/JSPS KAKENHI Grand Numbers JP 25220001, 18J01507.

References

[1] A. Casta˜no-Mart´ınez and F. L´opez-Bl´azquez. Distribution of a sum of weighted noncentral chi-square variables. TEST, 14(2):397–415, December 2005. [2] U. Graf. Introduction to Hyperfunctions and Their Integral Transforms. Birkh¨auser, Basel, 2010. [3] T. Hibi. Gr¨obner Bases: Statistics and Software System. Springer Japan, Tokyo, 2014. [4] I. Imai. Applied hyperfunction theory. Kluwer Academic Publishers, Dordrecht, 1992. [5] A. Kaneko. Introduction to Hyperfunctions. KTK Scientific Publishers, Tokyo, 1988. [6] T. Kawai. On the theory of fourier hyperfunctions and its applications to partial differential equations with constant coefficients. Journal of the Faculty of Science, the University of Tokyo, 17(3):467–517, December 1970. [7] T. Koyama and A. Takemura. Holonomic gradient method for distribution function of a weighted sum of noncentral chi-square random variables. Computational Statistics, 31(4):1645–1659, December 2016. [8] N. Marumo, T. Oaku, and A. Takemura. Properties of powers of functions satisfying second-order linear differential equations with applications to statistics. Japan Journal of Industrial and Applied Mathematics, 32:553–572, July 2015. [9] H. Nakayama, K. Nishiyama, M. Noro, K. Ohara, T. Sei, N. Takayama, and A. Takemura. Holo- nomic gradient descent and its application to the Fisher-Bingham integral. Advances in Applied Mathematics, 47:639–658, 2011.

13 [10] H. Ogata and H. Hirayama. Hyperfunction method for numerical integrations. Transactions of the Japan Society for Industrial and Applied Mathematics, 26(1):33–43, 2016. (in Japanese). [11] M. Sato. Theory of hyperfunctions. Sˆugaku, 10:1–27, 1958. in Japanese. [12] David Williams. Probability with Martingales. Cambridge University Press, Cambridge, 1991.

14